A Divide-and-Conquer System Based Radial Basis Function Network with its Algorithm of Maximizing Conditional Probability

This paper presents a divide-and-conquer system based radial basis function (DCRBF) network and its learning algorithm referred as maximizing conditional probability (MCP). This architecture is composed of several sub-RBF networks which have their input subspace. The output of DCRBF is a sum of the sub-networks´ outputs. We apply DCRBF to recurrent time series model. The experimental results have shown that the DCRBF outperforms the original RBF in the convergent speed and the generalization ability.